Short answer: AI is implemented in a business in five steps — measure where the time goes, pick one process that repeats often, write down the success criterion, build on tools you already have wherever possible, and train the team on what was built. The tool is chosen in step four, not step one.
Most conversations about AI in a company open with "which tool should we get?". That is understandable — tools are visible, they have a demo and a price on the website. But it is the question from the wrong end. A company that buys a tool before it knows where its time goes ends up with a tool that solves a problem it may not have.
This guide describes the order we use with clients. It is not the only possible one, but it has one advantage: every step produces something that stays useful even if the next one never happens.
01. Measure where the time goes
The first step is not technical. It is measurement: conversations with the people who actually do the work, an end-to-end map of the workflow, and an inventory of the systems data passes through — ERP, CRM, email, spreadsheets, files.
The aim is to tie every lost hour to a specific person, a specific step in their day, and an estimated value. "Sales loses time on quotes" is not a finding. "Two people in sales spend about an hour a day retyping the same data from an enquiry into the ERP and into the quote" is.
We call this step Diagnose, and it takes two to three weeks. The output is a workflow map, a systems map, and three to five automation opportunities with estimated ROI. The document is usable even if you do nothing after it — or continue the work with someone else.
02. Pick one process
Measurement usually produces more candidates than it makes sense to tackle at once. Choose the first process on three criteria:
- Frequency. Work done twenty times a day returns more time than work done once a month, even if the monthly task is more tedious.
- Predictability. If the outcome is usually the same — the same type of enquiry, the same document format, the same check — AI has leverage. If every case needs an experienced person's judgement, it is not a first candidate.
- Visibility to the people who do it. The first project should be something the team feels in their day. That builds the trust needed for the second.
Typical first candidates are dull: the same data typed into three systems, first drafts of replies to recurring questions, reconciling two reports that ought to agree. More examples are in Five workflows that pay for themselves in 90 days.
03. Write down the success criterion before you start
This is the step most often skipped, and the reason many AI projects never end as either a success or a failure — they just fade.
Before anything is built, write one sentence: what must be true after six weeks for the project to have worked. For example: "time from enquiry to sent quote drops from two days to same-day, for standard enquiries". Not "the team is happier" and not "the process is more modern".
A written criterion changes the conversation with leadership. On our Pilot it also has a financial consequence: if it is missed, the design and delivery fee is refunded in full.
04. Only now, choose the tool
Once you know which process you are changing and how success will be measured, the choice of tool narrows sharply — and is often cheaper than expected.
Each opportunity gets three questions: can it be solved with something the company already pays for (Claude, ChatGPT, Gemini, Microsoft 365, the existing ERP), is there an off-the-shelf product that solves it better, or does it need to be built? That build-vs-buy decision is the content of the phase we call Design, and it takes one week.
Two things are worth knowing in advance. First, the implementation price is not the only cost: there is also the monthly spend on AI models, which grows as the system gets used. Second, if data must not leave the company, the solution can run entirely on your own servers — more in AI security and compliance.
05. Build, hand over and train
The last step is building, testing and handover. A few things separate a handover that lasts from one that falls apart in three months:
- Evaluation before launch. A set of real examples used to check the system works, re-run after every change. More in Evals are the only feature you should ship before launch.
- An owner inside the company. Someone who knows how the system works and whom people go to when something looks wrong.
- Training on what was built, not a generic AI course. People adopt a tool when they see their own work in it.
After handover the system is yours: documented, and not dependent on a monthly contract to keep running.
The most common mistakes
- Buying company-wide licences before there is a concrete use case. The result is a tool used by a few enthusiasts and forgotten by everyone else.
- A pilot with no success criterion. After three months nobody can say whether it worked, so it never expands. Why this happens is covered in Why most AI pilots fail at month three.
- Starting with the most visible process instead of the most frequent one. A chatbot on the website is visible; triple data entry in the back office is expensive.
- A project with no owner inside the company. An outside partner can build the system, but cannot be the only one who knows how it works.
Frequently asked questions
How long does it take to implement AI in a business?
For one process, from measurement to a system running on your data, around six weeks is realistic. Measurement (Diagnose) takes two to three weeks, choosing the solution (Design) one week, and building one or two clearly defined solutions two to four weeks. A wider transformation is a series of such steps, not one large project.
How much does it cost to implement AI in a business?
It depends on the number of processes, the number of systems to connect, and where the solution runs after handover. We work to a fixed scope with the price agreed before work starts; how that price is formed is set out on AI implementation cost.
Can a small business implement AI?
Yes, and often faster than a large one, because it has fewer systems and a shorter path to a decision. The same order applies: measure where the time goes, then choose the tool. For a small business the measurement is shorter, but it should not be skipped.
Do we need our own AI team?
Not to start. You need one person in the company who owns the process being changed and stays accountable for the system after handover. An outside partner can do the technical build; ownership of the process cannot be outsourced.
What if our data cannot leave the company?
Then the solution runs on your infrastructure. The choice between cloud tools and a solution on your own servers is made during Design, based on which data enters the process and what GDPR and local data protection law say about it.